5th EAI International Conference on Big Data Innovation for Sustainable Cognitive Computing

Despite the intense improvements in existing architectures and the development of new deep-learning models, the core of the dense layers remains the same. Here, we are referring to the function used to process the input from the neurons of previous layers, also called the transformation function. In this study, we propose a new methodology for the computation of a neuron’s output, by adding a quadratic term to the conventional linear operation. With the aim of achieving a higher accuracy, we test the proposed process on a number of well-known data sets (MNIST, CIFAR). There is a significant improvement in the performance of a simple dense neural network. The initial accuracy itself is much higher when the new neurons are used. Initial convergence to higher accuracies is always much faster in the proposed model, and the computational time of the model is reduced to half (or even less). This enhancement can potentially be reflected in every deep learning architecture that uses a dense layer and will be remarkably higher in larger architectures that incorporate a very high number of parameters and output classes. The proposed model also fights issues like the vanishing gradient. Toward the end of the paper, we discuss the vast future scope of this study.

Decision-making is a critical task in our day-to-day applications. Any problem in this real world needs accurate solution at any point of time. There are several machine learning algorithms that can be adopted specifically for particular purpose and extreme learning machine (EML) is one among these algorithms. This paper presents a brief study of extreme learning machine and its combination with various metaheuristic algorithms, and its application to many problems specific domain. The performance achieved by these metaheuristic algorithms fused extreme learning machines are discussed in terms of relevant metrics. The details of the dataset used in the applications are also discussed.

Nanomaterials are used in almost every field of engineering. Synthesis techniques and conditions greatly affect the properties of synthesized nanomaterials. Identifying the nanomaterial from FESEM and TEM images with bare eyes is an exceedingly impossible task. Digital image processing techniques play a vigorous part in identifying the size and structure, and classifying them precisely helps scientists and investigators to use them in numerous applications. The advantages of digital image processing techniques increase the precision of object recognition in computer vision and pattern recognition. The proposed technique extracts various textural features such as kurtosis, skewness, and entropy from boron, iron, and silver nanoparticle images. The classification is done by using PNN and K-NN classifiers. The K-NN classifier has an accuracy of 80.00% for boron, 86.67% for iron, and 93.33% for the silver nanoparticle images, and the PNN classifier has an accuracy of 86.67% for boron, 93.33% for iron, and 93.33% for silver nanoparticle images. Hence, based on the experimentation, the proposed study suggested that the PNN classification with texture features is the best classifier used to classify the boron, iron, and silver nanoparticle images as compared to the K-NN classifier. Further, the results also are established manually with chemical experts, which proves the exhaustiveness of the proposed method.

Big data play an enormous role in the real-world industry, such as financial, banking, government sector, and healthcare, to store the enormous amount of data stored and processed. Here, big data analytics processing develops many applications but also leads to some challenges: data security, data structure, data standardization, and data storage and transfer. Big data are working more number of data to process, and nowadays, the stored data are increased exponentially in terabytes to zeta bytes, so we will not handle those increases by filtering each data to stored classified data value, removing the replicated data and the unnecessary data stored in big data. In this research, we involved big data analytics in machine learning technique (BDML) to reduce the challenges mentioned above using a classification algorithm. They are many types of classification algorithms that are there to predict the outcome of unstructured data. The primary purpose of the research is to reduce the rising size of data stored and sent, improve data security, and create organized data using machine learning techniques. A graphical model shows our aim of research attained with exemplary efficient implementation. In future, we will improve this technique to work with various real-world applications.

In the field of software engineering, Source Code Management is always challenging and difficult to cope in certain stages. Source Code Management (SCM) plays an important role in organizing, managing, and controlling the changes delivered by the developers to the source code, documents, and other entities. During the Software Development Life Cycle (SDLC) of a software product, many stakeholders of different teams also from different geographical locations are involved, including developers, testers, and analysts. With the involvement of a number of stakeholders, organizing, managing, and controlling the changes plays an important role. Software Configuration Management (SCM) is a process that helps methodically organize, control, and manage the changes in documents, code, and other entities during the SDLC of the software product. Currently, there is no system within LGSI that can track the changes made to the code by the various stakeholders involved. This study aims to overcome this problem; the LGSI proposed a new system called Change Tracker. The study finds that an individual developer can track the changes delivered to his/her interested files and creates permalinks delivered by LGSI developers then the permalinks will be reviewed by the review task team. In this paper, the model notifies the respective stakeholders when the changes are made and thus overcome the current system’s flaws.

As human beings yearn for better comfort and development in all aspects of life, they invent better technologies and adopt them for a better standard of living. This scenario prevails in the home automation field as well. The goal of the proposed system is to develop a sustainable green home with reduced energy and water wastage with improved security in all aspects. This chapter deals with the design and development of smart home management system incorporating mesh network and embedded system. We have developed a mesh-network-based smart home management system that is capable of controlling water levels, which is an advanced water tank management and control system, an advanced gas leakage management system that will automatically detect the leakage of gas and will actuate accordingly and advanced security system along with some sensing parameters, and home management system and garden management system. It provides an overview of the system architecture with insights into the mesh network formed using Raspberry Pi 4 and ESP32 with MQ Telemetry Transport (MQTT) protocol. The mesh network, layered architecture, and MQTT communication system are presented. The system is highly flexible, as newer nodes could be added according to user requirements. The scope of the system could be expanded beyond households into schools, hospitals, etc., to benefit a wider audience.

Due to the popularity of cloud services, it is unavoidable, that not just legitimate, but fraudulent registrations will happen. For a service with good reputation, it is essential to prevent fraud users. A common way is to filter these cases during the registration process by analysts. This chapter presents a novel decision support system that can recognize anomalous behavioral patterns and classify accounts based on the available data thus implementing an automated fraud prevention system. The process uses both supervised and unsupervised approaches, thus avoiding errors due to inaccurate labeling. As a supervised machine learning algorithm, random forest classifier and logistic regression are used, and as an unsupervised, auto encoder is used. The developed flow gives a recommendation to the analyst whether a new user is potentially fraud or not and provides feedback on the accuracy of analysts’ work based on the results of the unsupervised approach. The newly developed process is able to supervise the decisions made by analysts thus improving the labeling process. The main goal of this chapter is to present a new, more deterministic labeling workflow with the ability to provide feedback so it can improve the correctness of the training data set.

Data at present has emerged as the life blood for all businesses across the globe, as the world turns toward abundance of availability of data, it brings with it a few challenges as well; this paper deals with such challenges that the data scientists or analysts face will using big data models in the retail sector. The present research uses the DEA technique to analyze the variables that are efficient or inefficient while deploying the defined big data model in the retail sector. The findings of the study indicate that the variables like data security, cost of data have proved to be efficient during the model deployment, but the factors like data privacy, credibility, and the technology infrastructure proves to be inefficient during the deployment of the defined big data model in the retail sector.

In the recent days, the research area of handwritten character recognition has got much attention toward ancient inscriptions, since they contain lots of unfolding knowledge in the field of science, literature, astronomy, medicine, etc. The materials used to write these inscriptions are paper, palm leaf, stone rocks, and temple walls, etc., and these materials are now degrading in nature due to climatic conditions, ink bleeding, lack of attention, and unscientific storage. In this paper, the digitization and restoration of Kannada handwritten palm leaf with iterative global threshold based segmentation is developed. The performance estimation is measured by calculating MSE and PSNR values, and the image quality is compared to manual obtained results by epigraphists. The average values of PSNR and MSE are 6.198 and 0.234, respectively. The higher the PSNR and lower the MSE determines the quality of the image. The outcomes are also compared to other standard methods, namely, Souvola, Niblack, and Adaptive thresholding (Gaussion+Binary Inverse). The comparison studies confirm that the proposed algorithm is more effective than the other methods. The proposed algorithm is also implemented on the benchmark standard palm leaf dataset, i.e., the AMADI_LONTARSET.

Nowadays, massive data-intensive scientific applications are represented in the form of workflows. Workflows are the set of dependent or independent tasks groups that together execute the scientific applications effectively and quickly by acquiring computation resources from the various distributed environments, such as grid and cloud. On executing the dependent workflows, failure of a task leads to resubmission of the whole workflow again. It is proposed to predict the status of the tasks and execute them in the cloud resources. This is called as label tuning. Based on similarity checking and normalization mechanisms, tasks are categorized into Success/Partial Success and Failure/Partial Failure. Such multi-label classification is used to reschedule the tasks before the failure occurs. The experimental results show that predicting the task failure using multi-label classification provides 97.5% accuracy, which is 1.5% greater than binary label classification.

The aim of a supply chain team is to formulate a network layout that minimizes the total cost. In this research, the lowest production cost of the final product has been determined using a generalized plant location model. Furthermore, it is anticipated that units have been set up appropriately so that one unit of input from a source of supply results in one unit of output. The assignment problem is equivalent to distributing a job to the appropriate machine in order to meet customer demand. This study concentrates on reducing the cost of fulfilling the overall customer demand. Many studies have been conducted, and various algorithms have been proposed to achieve the best possible result. The purpose of this study is to present an appropriate model for exploring the solution to the assignment problem using the “Hungarian Method.” To find a feasible output of the assignment problem, this study conducted a detailed case study. The computational results indicate that the “Hungarian Method” provides an optimum solution for both balanced and unbalanced assignment problems. Moreover, decision-makers can use the study’s findings as a reference to mitigate production costs and adopt any sustainable market policy.

Technology cannot solve all age-related problems; however, they have the potential to relieve a significant amount of stress on elderly and their family. In this paper, existing memory assistive techniques for elderly are reviewed, and the idea of a smart memory companion is presented. The application leverages the personalized digital journal curated using historical and daily life experiences of the elderly to aid the aged in early stages of Alzheimer. The implementation proposed in this paper focuses on using automatic speech recognition and natural language processing techniques to build an intelligent conversational agent, retrieval engine, and an alerting module to help the elderly person in recalling and outdoor safe navigation. The smart application is meant to reduce the burden on the family to continuously monitor their aged loved ones and avoid caretakers unless necessary. The companion can be used to collect memories in visual and textual format from elderly, family, and caretakers to utilize it when needed. The companion can also be integrated with existing virtual assistants like Alexa and Google Home.

Emotion recognition is a technique for identifying and classifying human emotions by combining biosensing, face recognition, speech and voice recognition, machine learning, and pattern recognition. The electroencephalogram (EEG), which is based on brain signals, is used in the proposed work to identify the emotions. Here, the DEAP dataset is used to gather information about human emotional states, and machine learning algorithms like SVM, KNN, and MLP were used as the classifiers. In this proposed work, we used Power Spectrum Density as the Feature Extraction Technique and Welch’s Periodogram as the Signal Decomposition Technique. It was observed that gamma signals show more Valence, whereas Alpha signals show Arousal. To Cross validation system has been used to validate the performance of classifiers used in the work. Based on the locations of the electrode, EEG bands, and classifiers, accuracy comparisons were done. Out of the three classifiers, KNN scored more accuracy than other classifiers used.

COVID-19 pandemic is viewed as the most hazardous irresistible infection due to its quick spreading nature. Many a times while detecting the COVID using X-rays, pneumonia also got wrongly identified as COVID-19. Just to avoid this confusion, this paper proposes a Convolutional Neural Network for the classification of COVID-19 and pneumonia sicknesses. These illnesses damage the human lungs. Early analysis of patients contaminated by the infection can be treated and the patient’s life can be saved. We can also stop further spreading of the infection to other persons. The Convolutional Neutral Networks (CNN) model is deployed to help in the early detection of the infection using chest X-ray pictures, as it is one of the quickest and most easy approaches for diagnosing this infection. We propose CNN which uses MobileNet V2 Architecture. This proposed method classified the pneumonia and COVID with the accuracy of 98.6%.

It may be quite challenging to develop suitable techniques for analyzing the enormous amount of unstructured content of news items and extracting opinions from it. Understanding the attitudes, opinions, and feelings expressed in an online mention generally involves figuring out the emotional undertone of a string of words. Problems resulting in incoherent sentiment analysis are encountered in many of the existing approaches for sentiment analysis due to the presence of punctuation, ironic sentences, etc. Negative stopwords carry significant information about the sentiment of the sentence, but it is found that most of the approaches concerned with sentiment analysis remove these stopwords during the pre-processing stage. Thus, a suitable approach using negative stopwords has been proposed to minimize the information loss and improve the accuracy of the results obtained from sentiment analysis. The proposed approach has been evaluated using the dataset of various categories of news articles obtained from BBC, and it is found to yield an average accuracy of 85.75 for providing the sentiment polarity of various news items.

The financial industry is changing rapidly by offering different services with latest technologies to meet the expectations of the end users. Fintech is rapidly emerging in this digitization process, and financial companies are reaping huge benefits from applying the same. Current studies on this topic are limited, and further research is needed on this topic. This is the main motive behind this study and fixed its objective to find out the factors influencing the banking customers in adopting Fintech, by keeping Tamil Nadu as the research study location. This study is an analytical in nature in which banking customers are the respondents. A sample of 200 respondents is selected randomly and collected their opinion using a structured questionnaire. Tools like Mean, Standard Deviation, Factor Analysis, Kaiser-Meyer-Olkin (KMO) Measure, and Bartlett’s Test of Sphericity, Correlation, and Regression were used to analyze the collected data. Seventeen variables were constituted under three factors, namely, Conducive, Adaptability, and Security are used in this study to meet out the stated objectives and found that out of three factors, variables listed under Conducive factor has the significant impact on the Fintech adoption, whereas the other two factors are not significantly influencing the banking customers toward Fintech adoption.

The driver’s drowsiness detection system is the research conducted in the field of computer engineering to develop a system to prevent accidents caused by driver drowsiness. Due to the rise in traffic accidents, we now suffer a variety of losses that have an impact on both the economy of our country and the next generation of our nation. The rapid increase in the number of vehicles and their speed on the road has worsened traffic congestion and the likelihood of increased traffic accidents. In the upcoming years, it will be crucial to implement a smart accident prevention system because the number of fatalities is rising significantly. Drowsiness can occur in many ways, such as the feeling of being sleepy or a strong desire to sleep for an extended period. Therefore, it is crucial to assess the psychological and biological factors that may influence a driver’s reflexes and shorten reaction time. One of the biggest causes of vehicle accidents is fatigued or exhausted drivers. While operating a vehicle, such as an automobile, one must be concentrated, aware, and cautious. This study presents a method that combines the Internet of Things (IoT) and a physiological approach to assessing blood oxygen levels to identify drivers who are drowsy while operating an automobile. Our technique, which detects tiredness in the driver’s blood in real-time by measuring blood oxygen levels, is effective. Drowsy driving has been criticized for a significant number of traffic accidents in recent years. As a result, we came up with the idea of detecting drowsiness while driving using a pulse oximetry sensor (SpO2). Next, the level of blood oxygen in the driver’s blood is calculated to determine and evaluate the driver’s level of drowsiness. Even though there are numerous other sensors, the goal of utilizing SpO2 is due to the accuracy of the results obtained. The result can be achieved by observing the change in oxygen levels caused by drowsiness. We utilize the SpO2 sensor in our research and projects since the oxygen level predicts the outcome. The alarm system that comes with drowsiness detection will alert the driver due to their lack of concentration and drowsiness. To avoid an accident, the alarm notifies the motorist that they are now exhausted and suggests having a break. Currently, road networks are a major part of human life, so they must be fixed. This research paper is very cost-efficient, and the outcomes obtained are accurate by using the SpO2 sensor for detecting the driver’s drowsiness and the alarm helps the driver in regaining consciousness. Thus, there will be a decline in the rate of road accidents compared to the past. This system additionally uses Arduino UNO & MAX 30100 pulse oximetry sensor as the processing unit and has an LCD for displaying the status.

Investment authorities are broad financial institutions that carefully manage investments on behalf of the national government using a long-term value development approach. To provide a stronger structure or framework for Investment Authorities to govern the distribution of funds to public and private markets, we have started research to create a blockchain-based prototype for managing and tracking numerous finances of such authorities. We have taken the case study of Oman Investment Authority (OIA) of Sultanate of Oman. Oman’s wealth is held in OIA. It is the organization that oversees and utilizes the additional capital generated by oil and gas profits in public and private markets. Unlike other Omani funds, this one focuses primarily on assets outside the Sultanate. The operation of the OIA entails a large number of transactions, necessitating a high level of transparency and administration among the parties involved. Currently, OIA relies on various manuals to achieve its goals, such as the Authorities and Responsibilities manual, the Investment Manual, and the Code of Business Conduct, among others. In this paper, we propose a blockchain-based framework to manage the operations of OIA. Blockchain is a part of the Fourth Industrial Revolution, and it is reshaping every industry. The main components of every blockchain are assets and participants. The funds are the major assets in the proposed study, and the participants are the various fund shareholders/recipients. The blockchain’s transactions are all safe, secure, and immutable, and it is part of a trust less network. The transactions are simple to follow and verify. By replacing intermediary firms with smart contracts, blockchain-based solutions eliminate any middlemen in the fund allocation process.